SpectralUnmix: A Torch-Based Regularized Non-negative Matrix Factorization

TL;DR

SpectralUnmix is a Torch-based R package for regularized non-negative matrix factorization, enabling spectral data analysis with GPU support and smoothness regularization.

Contribution

It introduces a new R package implementing regularized NMF with proximal-gradient updates and GPU acceleration, tailored for spectral data analysis.

Findings

Successfully applied to stellar spectra data

Recovered NMF components comparable to principal components

Demonstrated GPU acceleration benefits

Abstract

We present SpectralUnmix, an R package for regularized non-negative matrix factorization (NMF), implemented in torch with optional GPU acceleration. The package estimates low-rank non-negative representations through proximal-gradient updates and allows smoothness regularization along the spectral axis. As a compact demonstration, we apply the method to a subset of stellar spectra and compare the recovered NMF components with principal-component directions and representative stellar spectra. The package is released under the MIT license at \href{https://rafaelsdesouza.github.io/SpectralUnmix/}{this repository}.